feat(prompt): v6 auditor — journal legitimacy (H10) and scoring formula fix

[lentago-claude-runner]

Origin

Chris resolved the #7 data-source question in favor of a hybrid indexing-plus-free-blacklist approach and chose to bundle the #43 scoring-formula fix into the same v6 release, since both are vNext prompt changes to the same new file. Both issues close together here.

Summary

• Heuristic 10 — journal legitimacy and predatory-venue flagging (closes [v4] Implement predatory journal flagging #7): hybrid whitelist-plus-community-list approach. DOAJ, PubMed/MEDLINE, Scopus, and Web of Science serve as primary positive signals; Beall's archived list and Stop Predatory Journals provide secondary corroboration only. Flags Elevated in isolation; escalates to High when combined with any other heuristic trigger. Factual, non-accusatory language throughout — "predatory" is never used as a verdict; "potentially predatory" or "unverified venue" at most. Grey literature is exempt.
• Scoring formula fix (closes Scoring formula: D×3 base cost makes the headline score misleading for large reference lists #43): removed the D × 3 base cost. New formula: Score = 100 − (H × 12) − (E × 5) − (M × 2), floored at 0. A fully-clean reference list of any length now scores 100. % Defensible is surfaced as a prominent complementary signal in the Executive Dashboard. Risk-band thresholds and gauge color zones recalibrated (90–100 green / 70–89 amber / 50–69 orange / 0–49 red).

Files changed

File	What changed
prompts/v6-auditor.md	New file — copy of v5 with both changes integrated
test-sets/predatory-venues.md	New file — 5-reference H10 test set (3 traps, 2 clean controls)
docs/heuristics.md	Full H10 section; scoring formula section; v6 version history row; H10 in COPE table and heuristic interaction
README.md	Ten heuristics; updated scoring formula; Shipped in v6 roadmap section; updated repo layout and Running an Audit reference
CLAUDE.md	v6 as current; v5 as previous live version for diffing; H10 in the heuristics description
roadmap/v4-features.md	Predatory journal flagging marked done (shipped in v6); prioritized list updated
test-sets/real-articles/README.md	Scoring note updated to reflect v6 formula; v5 scores noted as not comparable

What is NOT in this PR

• v5 is frozen: prompts/v5-auditor.md, test-sets/adversarial-30.md, and all committed reports are unchanged.
• No audit run. See validation note below.

Validation follow-up (suggested)

v6 needs a fresh baseline run before being treated as production-validated. The baseline should confirm:

1. H10 detection: test-sets/predatory-venues.md — all three trap entries flag correctly (two Elevated for H10 alone, one High for H10 + H7), and both clean controls (JOGNN, BMC Pregnancy and Childbirth) remain Defensible.
2. Scoring calibration: clean real articles (Ahmadinezhad, Patriksson) should score near 100 under v6 rather than the 5 and 22 they scored under v5; the dirty corpus (Amarnath, Madhukar) should still score low.
3. Adversarial-30 regression: v6 should match or exceed v5's adversarial-30 sensitivity (100%) and specificity (92.9%). Because the formula changed, the numerical scores will differ, but the H/E/M/D tier assignments should be stable.

Because the scoring formula changed, v6 baseline metrics will not be directly comparable to the v5 numbers committed in reports/adversarial-30-v5-2026-06-19.html.

Prompt-Origin

Implemented by claude-runner[bot] per task prompt. Origin context: Chris resolved the #7 data-source question in favor of a hybrid indexing-plus-free-blacklist approach and chose to bundle the #43 scoring-formula fix into the same v6 release since both are vNext prompt changes to the same new file.

Closes #7
Closes #43

[cpitzi]

Reviewer correction pushed before merge. During review I verified the H10 clean-control references against Crossref and found both had broken DOIs that would have made them flag under other heuristics, defeating their purpose as clean controls:

• Ref 2: 10.1016/j.jogn.2021.06.007 resolved to Shin et al. (a different paper), not the cited Edmonds et al. → replaced with Faridvand et al. 2018, Int. J. Nursing Practice, 10.1111/ijn.12659 (verified resolves to the cited paper).
• Ref 5: 10.1186/s12884-015-0448-7 → 404; the Vogel Robson paper is actually in The Lancet Global Health, not BMC. Corrected journal + DOI to 10.1016/S2214-109X(15)70094-X (verified).

Both replacements re-verified HTTP 200 and resolve to the cited paper. Also bumped CLAUDE.md's heuristic count nine→ten to match the README. The H10 prompt logic and the #43 scoring fix were verified correct and unchanged.